HYDROLOGICAL PROCESSES

Hydrol. Process. (2016)

Published online in Wiley Online Library
(wileyonlinelibrary.com) DOI: 10.1002/hyp.10789

Error distribution modelling of satellite soil moisture

measurements for hydrological applications
Lu Zhuo,1* Qiang Dai,2 Tanvir Islam3,4 and Dawei Han1
1
WEMRC, Department of Civil Engineering, University of Bristol, Bristol, UK
2
Key Laboratory of Virtual Geographic Environment of Ministry of Education, School of Geography Science, Nanjing Normal University, Nanjing, China
3
NASA Jet Propulsion Laboratory, Pasadena, CA, USA
4
California Institute of Technology, Pasadena, CA, USA

Abstract:
Satellite-based soil moisture data accuracies are of important concerns by hydrologists because they could significantly influence
hydrological modelling uncertainty. Without proper quantification of their uncertainties, it is difficult to optimize the
hydrological modelling system and make robust decisions. Currently, the satellite soil moisture data uncertainty has been limited
to summary statistics with the validations mainly from the in situ measurements. This study attempts to build the first error
distribution model with additional higher-order uncertainty modelling for satellite soil moisture observations. The methodology
is demonstrated by a case study using the Soil Moisture and Ocean Salinity satellite soil moisture observations. The validation is
based on soil moisture estimates from hydrological modelling, which is more relevant to the intended data use than the in situ
measurements. Four probability distributions have been explored to find suitable error distribution curves using the statistical
tests and bootstrapping resampling technique. General extreme value is identified as the most suitable one among all the curves.
The error distribution model is still in its infant stage, which ignores spatial and temporal correlations, and nonstationarity.
Further improvements should be carried out by the hydrological community by expanding the methodology to a wide range of
satellite soil moisture data using different hydrological models. Copyright © 2016 John Wiley & Sons, Ltd.

KEY WORDS error distribution model; uncertainty modelling; satellite soil moisture; hydrological modelling; Soil Moisture and
Ocean Salinity (SMOS); Xinanjiang (XAJ)
Received 25 June 2015; Accepted 8 January 2016

INTRODUCTION and hence are not able to provide sufficient information at a

catchment scale (Al-Shrafany et al., 2013; Srivastava et al.,
Soil moisture is crucial to the improvement of most water-
2013b). Another technique to obtain catchment-scale soil
related systems, including real-time hydrological modelling
moisture relies on land surface modelling (e.g. Noah land
at catchment scales (Ottlé and Vidal-Madjar, 1994; Merlin
surface model with multiparameterization options (Noah-
et al., 2006; Brocca et al., 2010; Guingla et al., 2012). This is
MP) and Variable Infiltration Capacity (VIC) models) or
because real-time hydrological modelling suffers from the
utilizing data assimilation to integrate land surface models
accumulation of errors in evapotranspiration and soil
with ground measurements. However, model-based soil
moisture. This shortcoming can often result in poor
moisture estimation tends to suffer from time drifts
performance of flow prediction after some dry periods
problem (e.g. error accumulation over times). Alternative-
because of the inaccurate calculation of the antecedent soil
ly, remote sensing is a feasible way to monitor catchment-
moisture (Ottlé and Vidal-Madjar, 1994; Brocca et al., 2008).
scale soil moisture information because it provides
The in situ measurements, remote sensing and land
integrated estimation over pixel scales (Engman and
surface models are the main sources of soil moisture
Chauhan, 1995). Optical and thermal infrared remote
estimation for hydrological applications. Soil moisture
sensing (Hulley et al., 2010; Hain et al., 2011), as well as
has a high variability in both space and time, so it is
passive and active microwave remote sensing techniques
difficult to estimate accurately (Walker et al., 2004; Wang
(Kerr et al., 2010; Sánchez-Ruiz et al., 2014), has been
and Qu, 2009). The majority of the available ground-
extensively used for soil moisture monitoring over various
based soil moisture observations are sparsely distributed
conditions of topography and vegetation.
These data products have the potential to be used in
*Correspondence to: Lu Zhuo, Water and Environmental Management real-time flood forecasting systems. However, those
Research Centre, Department of Civil Engineering, University of Bristol,
Bristol, UK. measurements can be affected by several error sources
E-mail: lu.zhuo@bristol.ac.uk (e.g. algorithms, sensors and physical processes)

Copyright © 2016 John Wiley & Sons, Ltd.

 L. ZHUO ET AL.
(Dorigo et al., 2010). Quantification of such uncertainties
is particularly important for applying the soil moisture
datasets in real-time flood forecasting systems (Brocca
et al., 2010). More importantly, this is the foundation to the
optimal modelling performance in using such soil moisture
datasets. Because satellite soil moisture data are global
and widely available, they are of great interest to the
hydrological community. Although there are many
studies on exploring the uncertainty of satellite soil
moisture estimates in hydrological applications, they are
mainly represented as summary statistics [such as root
mean square error, correlation or Nash–Sutcliffe
efficiency (NSE)] (Panciera et al., 2008; Albergel et al.,
2011, 2012; Piles et al., 2011; Al Bitar et al., 2012;
Collow et al., 2012; Lacava et al., 2012b; Srivastava
et al., 2013a,b, 2014), and there is a lack of attention on
the error distribution model (such as probability density
function, spatial and temporal correlation, and
nonstationarity).
It is recognized that because of the complexity of soil
moisture in the field, no reference dataset can provide the
absolute truth (Wagner, 2008), so it is important that the
assessment of soil moisture products should be based on
different benchmarks (such as in situ measurements,
modelled soil moisture data from land surface model or
hydrological model and different sensors) according to
their applications. Conventionally, satellite soil moisture
data have been mainly validated using in situ measure-
ments, which are useful in many applications, but it
should not be considered as the only benchmark. For
example, Dorigo et al. (2010) used three independent
satellite soil moisture datasets to derive their error
characterizations (albeit only summary statistics such as
root mean square error are derived) without using ground-
based estimates (the so-called ‘triple collocation error
estimation’). The reason is explained by the authors as
‘a spatially coherent assessment of the quality of the
various globally available datasets is often hampered
by the limited availability over space and time of
reliable in-situ measurements’. As Wagner (2008) has
explained that ‘there is no universal remote sensing
method (= sensor + algorithm) that satisfies all user
requirements’, an excellent performance of the soil
moisture data in one application field may have a poor
performance in another field. For example, in climate
change modelling, the global climate models usually have
spatial resolutions in hundreds of kilometres, so a soil
moisture product suitable for such applications may not be
optimal for studies in crop growth monitoring in
agriculture, and vice versa. Because this study aims at
hydrological modelling applications, the benchmark is
based on the effective catchment soil moisture data
estimated from a well-known operational hydrological
model Xinanjiang (XAJ).
Copyright © 2016 John Wiley & Sons, Ltd.
In contrast to previous studies that have only focused
on summary statistics, in this study, we aim to build the
first error distribution model of satellite soil moisture
estimates. The Soil Moisture and Ocean Salinity (SMOS)
level-3 soil moisture product is used as a case study (Kerr
et al., 2001). SMOS satellite is chosen because it is the
first mission dedicated to monitoring direct surface soil
moisture on a global scale (Kerr et al., 2010) and owns a
higher soil penetrating capability (~5 cm of soil surface at
1.4 GHz) because of its longer wavelength (L-band).
Another dedicated soil moisture satellite called the Soil
Moisture Active/Passive mission (1.20–1.41 GHz;
Entekhabi et al., 2010) was only launched in early
2015, which is not suitable because of its short record.
Long-term satellite and model datasets are processed (i.e.
4 consecutive years). Four traditional probability distri-
butions are introduced to identify the optimal error
distribution model(s) to describe the features of the
SMOS soil moisture uncertainty, by employing the
statistical tests and bootstrapping resampling technique.
The optional error distribution models are then quantified
further by second-order error distribution modelling.
DATA, CATCHMENT AND METHODOLOGIES
Study area and datasets
The chosen catchment (Pontiac, 1500 km2) is a part of
the Vermilion River situated in the central Illinois of the
USA (40.878°N, 88.636°W). This region is influenced
primarily by hot summer continental climate (Peel et al.,
2007) and covered mainly by cropland (Hansen et al.,
1998; Bartholomé and Belward, 2005) on Mollisols soil
type (Webb et al., 2000). The average altitude of the
catchment is 188 m above mean sea level, and its average
annual rainfall is 867 mm. The layout of the Pontiac
catchment is shown in Figure 1 along with the location of
its flow gauge, the National Aeronautics and Space
Administration Land Data Assimilation Systems 2
(NLDAS-2) grids and distribution of the river network.
The NLDAS-2 (Mitchell et al., 2004) precipitation (P)
and potential evapotranspiration (PET) at 0.125° spatial
resolution and daily temporal resolution (converted from
hourly resolution) are used to drive the XAJ model. More
detail about the NLDAS-2 datasets can be found in Xia
et al. (2012) and Zhuo et al. (2015b). Only the NLDAS-2
grid points within the catchment area are selected. As a
result, there are a total of 20 grid points covering the
whole area. Furthermore, the selected PET and P datasets
have been converted into one catchment-scale dataset
using the weighted average method for the usage in the
lumped XAJ model. The US Geological Survey daily
flow data from January 2010 to April 2011 have been
chosen for the calibration of the XAJ model, and the
Hydrol. Process. (2016)
 ERROR DISTRIBUTION MODELLING OF SATELLITE SOIL MOISTURE MEASUREMENTS
Figure 1. Location of the Pontiac catchment with the flow gauge and National Aeronautics and Space Administration Land Data Assimilation Systems 2
(NLDAS-2) grid points over the river network
period of May 2011 to December 2011 is used for the
validation. The 2-year data restriction is due to the
unavailability of the flow data in the catchment. However,
the short data length can still serve the purpose of this study
because of the following reason. The beauty of the XAJ
model is that its calibration can be based on short data
period. A study on calibration data selection of a XAJ-like
model (i.e. Probability Distributed Model (PDM)) has
shown that even a 6-month data can be sufficient for
excellent model calibration (Liu and Han, 2010). The
reason is that the information content in the calibration data
is a key factor instead of the length of the data period. As
long as there is enough perturbation to model parameters, it
is not necessary to have a much longer calibration time
period. We have found that the 1-year calibration data have
sufficient information to properly derive the effective
model parameters, and the validation result has also
demonstrated the effectiveness of the model. The level-3
SMOS soil moisture dataset is obtained from the SMOS
Barcelona Expert Centre (SMOS-BEC, http://cp34-bec.
cmima.csic.es) and has been processed for the period from
January 2010 to December 2013. In addition, the soil
moisture dataset under frozen condition has been removed
(i.e. air temperature below 0). This is because SMOS is
incapable of retrieving good quality soil moisture mea-
surements under frozen weather (Lacava et al., 2012a;
Al-Yaari et al., 2014; Wagner et al., 2014; Zhuo and Han,
2015). It has been further averaged into one catchment-
scale dataset by applying the weighted average method.
The MATLAB code has been used for all the modelling
works carried out in this study.
Uncertainty and its quantification methods
In a hydrological model, the soil moisture input can be
described by ensembles with stochastic elements, and the
usage of error distribution modelling allows us to better
understand the system (Jain and Singh, 2003; Beven,
2006). By analysing the error distribution models of the
input dataset, a decision can be made based on a range of
possible outcomes instead of a fixed dataset; this is rather
Copyright © 2016 John Wiley & Sons, Ltd.
important in water resources management (Jain and
Singh, 2003).
As previously mentioned, there exist many error
sources in satellite soil moisture measurements. However,
in real life, it is extremely difficult to investigate these
error sources separately and discover their interdepen-
dencies. The statistical error distribution model of satellite
soil moisture established with the assistance of hydrolog-
ical model could be a practical and efficient solution. The
error distribution model may change with time (i.e.
nonstationarity), but it is more challenging to build such a
model from the start. In this study, the model will be built
by exploring the significant characteristics of satellite soil
moisture errors from the long-time historical records
without considering the nonstationarity. One of the
simplest empirically based error distribution modelling
is via the systematic error analysis. It collects the
information of differences or ratios between satellite-
estimated and hydrologically modelled soil moisture
pairs. This methodology has been proven to be effective
and time efficient in numerous studies (Smith and
Krajewski, 1991; Anagnostou et al., 1998; Seo et al.,
1999; Borga and Tonelli, 2000) and provides a
fundamental framework for more complex error distribu-
tion models (Dai et al., 2014b).
In quantifying the uncertainty of SMOS soil moisture
estimates, it is assumed that the ‘true’ soil moisture (ST)
for assessing the observed satellite soil moisture (SS) is
based on the XAJ soil moisture simulations (SX), which is
used as the hydrological benchmark. The ‘true’ soil
moisture can be described by the following equation:
ST ¼ fSX jSS g (1)
In another word, the probable true soil moisture
corresponds to a realization of the distribution of the
XAJ soil moistures depending on the given satellite
estimates. Based on this, the satellite soil moisture
uncertainty can be shown as
Hydrol. Process. (2016)
 L. ZHUO ET AL.
ST ¼ M ðSS Þ þ εðSS Þ (2)
where M(SS) stands for the systematic errors of satellite
estimates and ε(SS) represents the random error compo-
nent.
In this study, it is mainly focused on quantifying the
total error observed when comparing SMOS and XAJ soil
moisture values and optimizing it with potential theoret-
ical models, so that the main parameters and statistical
characteristics of the observed error distribution model
can be derived. Four commonly used probability
distributions are explored to represent the error distribu-
tion of the SMOS soil moisture estimates.
The SMOS soil moisture product
The SMOS satellite was launched at the end of 2009
and has been providing soil moisture data for more than
5 years. SMOS acquires brightness temperature at the
frequency of 1.4 GHz [near surface soil moisture
(approximately 5 cm)]. The spatial resolution of the
SMOS products is 35–50 km (Kerr et al., 2001, 2010)
with soil moisture retrieval unit in m3/m3. SMOS offers a
global coverage at the equator crossing the times of 6:00 h
(local solar time, ascending) and 18:00 h (local solar time,
descending) (Kerr et al., 2012).
Soil Moisture and Ocean Salinity Barcelona Expert
Centre provides soil moisture measurements at various
temporal resolutions: daily, 3 days, 9 days, monthly and
annually. The BEC soil moisture products come in two
formats: the Icosahedral Snyder Equal Area projection
with aperture 4, resolution 9 grid with its shape of cells as
hexagon (Pinori et al., 2008) and the Equal Area Scalable
Earth grid with a spatial resolution of ~25 km × 25 km. In
this study, the daily soil moisture dataset with the Equal
Area Scalable Earth grid is chosen, because it is more
widely used. The main algorithms in generating the BEC
level-3 surface soil moisture are the same as the one
utilized by the European Space Agency for retrieving the
standard level-2 soil moisture products (Kerr et al., 2012).
In this study, only the descending SMOS data are used,
because we have found that the performance of
descending retrievals is much better than ascending’s in
this catchment (Zhuo et al., 2015a; Zhuo and Han, 2015).
This could be partly explained by the fact that this area is
highly affected by radio frequency interference, which
preferentially affects the ascending retrievals because of
the SMOS antenna pattern (Collow et al., 2012).
The XAJ hydrological model
There are many hydrological models available global-
ly, and in this study, a widely used model called XAJ is
employed. It is a fairly general conceptual lumped
rainfall–run-off model, which is capable of simulating
Copyright © 2016 John Wiley & Sons, Ltd.
flows at a variety of catchment scales over the globe
(Wang, 1991; Zhao, 1992; Khan, 1993; Zhao et al., 1995;
Zhuo et al., 2014). Moreover, its model concept and
structure are representatives to many typical conceptual
hydrological models. It is able to account for soil water
content in the system with a suitable time step and easy
access of data inputs (i.e. P and PET only) (Peng et al.,
2002). In hydrology, soil moisture deficit (SMD) or
depletion is a significant indicator of soil water content,
which stands for the amount of water to be added to a soil
profile to bring it to the field capacity (Calder et al., 1983;
Rushton et al., 2006). It has been shown by considerable
number of studies that the three-layer XAJ model is very
effective in generating SMD from the hydrological
forcing (Zhao, 1992; Zhao et al., 1995; Zhuo et al.,
2014). The main concept of XAJ is the run-off generation
on repletion of storage, which indicates that run-off does
not emerge until the soil water content of its aeration zone
reaches the field capacity. As shown in Figure 2, the
structure of the XAJ model comprises an evapotranspi-
ration module, a run-off production module and a run-off
routing module. The XAJ model has 16 parameters that
are calibrated in this study, by finding the optimal
performance in flow simulation (Zhao, 1980, 1992). The
three-layer SMDs are estimated to determine the effect of
drying and wetting on the catchment soil storage, and in
this study, only the SMD from the surface layer is used
because it is more scale matched with the satellite surface
soil moisture retrieval.
Bootstrapping resampling
Bootstrapping is a random sampling technique firstly
presented in Efron (1979), and a very informative and
readable account is given by Efron and Tibshirani (1986).
It has been used widely in systematic and clinical studies
(Sauerbrei and Schumacher, 1992; Hillis and Bull, 1993).
The reason for using bootstrapping is that it is able to
provide an indirect way of assessing the characteristics of
the distribution underlying the sample data (Adèr and
Mellenbergh, 2008). Moreover, bootstrapping is capable
of controlling and checking the stability of the results, yet
it only requires simple computing procedures. The
general idea of bootstrapping is the usage of ‘sampling
with replacement’, and more details regarding
bootstrapping can be found in Efron (1979).
Performance indicators
Nash–Sutcliffe efficiency (Nash and Sutcliffe, 1970) is
used to assess the performance of the XAJ model as well
as to select the suitable error distribution curves during
bootstrapping resampling process. NSE is the most
common and important performance measure used in
hydrology, and it is calculated using the following equation:
Hydrol. Process. (2016)
 ERROR DISTRIBUTION MODELLING OF SATELLITE SOIL MOISTURE MEASUREMENTS

Figure 2. Flowchart of the Xinanjiang model (Zhao, 1992), where W is the areal mean tension water storage with three components WU, WL and WD
representing upper, lower and deep soil layers, respectively; S is the areal mean free water storage; Fr is the factor of run-off producing area related to W;
IMP is the factor of impervious area in a catchment; RB is the direct run-off produced from the small portion of impervious area; R is the total run-off
generated from the model with surface run-off (RS), interflow (RI) and groundwater run-off (RG) components. These three run-off components are then
transferred into QS, QI and QG and combined as the total sub-catchment inflow (T) to the channel network. The flow outputs Q from each sub-catchment
are then routed to the catchment outlet to produce the final flow result (TQ)

∑ni¼1 ðyi  xi Þ2
NSE ¼ 1 
∑ni¼1 ðxi  xÞ2
where xi is the observed values and yi is the simulated
values. n is the number of data pairs.
Chi-squared test is a statistical test commonly used to
compare the observed data [e.g. cumulative distribution
function (CDF) of the observed error] with the data that
would be expected to obtain according to a specific
hypothesis (Mantel, 1963). The formula for calculating
chi-square (χ 2) is
ðxi  yi Þ2
χ ¼
2
∑ni¼1
yi
Pearson product moment correlation coefficient (r) is
used to evaluate the linear relationship between two
variables, which is defined as
nð∑xi yi Þ  ð∑xi Þð∑yi Þ
r ¼ rhffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ih
n∑x2i  ð∑xi Þ2 n∑y2i  ð∑yi Þ2
Spearman rank correlation coefficient (rsp) is a
nonparametric method for evaluating the degree of
correlation between two independent variables (Chen
et al., 2013; Dai et al., 2014a). Because of its ability to
cope with nonlinear correlation as well as linear
correlation, it is used in addition to r to check the
existence of nonlinearity between two sets of data. Its
formula is
n
6∑ d 2i
(3)
r sp ¼ 1  i¼1
(6)
n3  n
where di is the difference between ranks for each data pair
(xi, yi).
RESULTS
SMD estimated from XAJ model
First, the surface SMD is simulated by driving the XAJ
model. When running the XAJ model, the initial two
(4)
thirds of the data from the period of January 2010 to
December 2011 is used for model calibration while the
rest one third of the data is utilized for validation; this is a
common procedure used in hydrological modelling.
Because the XAJ model has been proven as quite reliable
(Zhao, 1992; Zhao et al., 1995), this allows us to
i (5) confidently apply the model for the periods different from
those used for calibration. The calibration procedure
focuses especially on the modelling of three-layer SMDs
and the distribution of the total run-off (e.g. surface run-
off, interflow and groundwater), as well as a good
agreement between the modelled flow and measured flow.
The performance of the model is judged by the NSE
indicator as its objective function. The optimal values of
the XAJ parameters, along with their initialization values
used in this study, are shown in Table I. The overall
performance shows an NSE value of 0.806 for the
calibration and 0.804 during the validation, which reveals

Copyright © 2016 John Wiley & Sons, Ltd. Hydrol. Process. (2016)
 L. ZHUO ET AL.

Table I. The Xinanjiang model parameters used in the Pontiac catchment

Symbol Model parameters Unit Optimal value Range

SM Areal mean free water capacity of the surface soil layer, which mm 31.48 10–50
represents the maximum possible deficit of free water storage
KG Outflow coefficients of the free water storage to groundwater relationships — 0.10 0.10–0.70
KSS Outflow coefficients of the free water storage to interflow relationships — 0.19 0.10–0.70
KKG Recession constants of the groundwater storage — 0.31 0.01–0.99
KKSS Recession constants of the lower interflow storage — 0.01 0.01–0.99
CS Recession constant in the lag and route method for routing through — 0.26 0.10–0.70
the channel system with each sub-catchment
WUM Averaged soil moisture storage capacity of the upper layer mm 46.96 30–50
WLM Averaged soil moisture storage capacity of the lower layer mm 39.08 20–150
WDM Averaged soil moisture storage capacity of the deep layer mm 30.11 30–400
IMP Percentage of impervious and saturated areas in the catchment % 0.00 0.00–0.10
B Exponential parameter with a single parabolic curve, which represents — 0.70 0.10–0.90
the non-uniformity of the spatial distribution of the soil moisture
storage capacity over the catchment
C Coefficient of the deep layer that depends on the proportion of the — 0.49 0.10–0.70
catchment area covered by vegetation with deep roots
EX Exponent of the free water capacity curve influencing the development — 1.93 1.10–2.00
of the saturated area
L Lag in time — 0.00 0.00–6.00
V Parameter of the Muskingum method m/s 0.43 0.40–1.20
dX Parameter of the Muskingum method — 0.18 0.00–0.40

Figure 3. Time series of rainfall, measured and simulated flows for the Pontiac catchment, during (a) calibration and (b) validation. OBS stands for the
flow observations.

Copyright © 2016 John Wiley & Sons, Ltd. Hydrol. Process. (2016)
 ERROR DISTRIBUTION MODELLING OF SATELLITE SOIL MOISTURE MEASUREMENTS

that the model works very well. The time series plots of
rainfall and flow during the calibration and validation
periods are illustrated in Figure 3. Further details on
calibration and validation of the XAJ model in this
catchment are discussed by Zhuo et al. (2015a).
SMD derived from SMOS soil moisture
In order to model the error distribution of the SMOS
soil moisture estimates for hydrological applications, the
4-year SMOS volumetric soil moisture dataset needs to be
converted into the hydrological SMD initially. First, the
correlation between the SMOS volumetric soil moisture
and the XAJ SMD is studied. As shown in Figure 4a, the
Spearman rsp and Pearson r correlation statistics between
the SMOS soil moisture estimates and XAJ SMD are
calculated, which yield almost the same values, indicating
that there is no strong nonlinearity. For this reason, a
linear fitting method is used for the derivation of SMOS
SMD. As seen in Figure 4a, the upper part of the XAJ
SMD represents the wilting point, while the lower range
represents the field capacity point of the soil. The values
depend on the soil properties and vegetation cover. The
comparison result between the derived SMOS SMD and
the XAJ SMD is presented in Figure 4b, with NSE
calculated as 0.492. It can be seen that there is an
overestimation of the SMOS SMD when surface soil is
wetter (i.e. smaller XAJ SMD) and a slight underestima-
tion when surface soil dries out (i.e. larger XAJ SMD).
This observed error (i.e. the differences between the XAJ
SMD and the SMOS SMD) is then investigated further by
first-order and second-order uncertainty modelling. For
the XAJ SMD, there is also an issue with the maximum
ceiling at 0.047 where many points are concentrated. We
will explore it in the discussion.

Figure 4. Soil Moisture and Ocean Salinity (SMOS) soil moisture deficit (SMD) derived from its volumetric soil moisture dataset, based on Xinanjiang
(XAJ) SMD: (a) a linear relationship between the XAJ SMD and the SMOS volumetric soil moisture estimates; (b) the derived SMOS SMD and a
comparison against the XAJ SMD

Copyright © 2016 John Wiley & Sons, Ltd. Hydrol. Process. (2016)
 L. ZHUO ET AL.

First-order error distribution modelling

This section is carried out to find the most represen-
tative curves to describe the observed discrepancy
between the SMOS SMD and the XAJ SMD. Here, the
observed discrepancy is fit with four well-known
distributions in hydrology: Gaussian, extreme value
(EV), general extreme value (GEV) and logistic distribu-
tions. Distribution formulas together with their parameters
are shown in Table II. More details about these four
distributions can be found in Walck (2007). It can be seen
that almost all the selected distributions are formed by
two parameters: a scale parameter and a location
parameter, except the GEV. The GEV introduces an
additional parameter called the shape parameter, which is
able to further describe the tail behaviour of a distribution.
The fitted results of the four distributions can be
observed from their histograms and CDFs plotted in
Figure 5. The skewness of the observed error is calculated
as 0.339, so it is right tailed. In Figure 5a, it is clear to see
that EV is more left skewed than others. For the
simulation of peak occurrence, the logistic shows the
highest value, while the Gaussian and GEV give lower
peaks (but very close to each other). The CDF plots are
also presented in Figure 5b. It shows that the Gaussian,
Figure 5. (a) Histograms of the four distributions and the histogram plot of
logistic and GEV curves are all very close to the the observed error; (b) CDF plots of the four distributions and the
observation; however, it is difficult to judge which one observed error. CDF, cumulative distribution function; EV, extreme value;
is the best based on visual inspection. Therefore, the chi- GEV, general extreme value
squared test is implemented, and the results are shown in
Table III. It is noted that a 5% level of statistical
significance is adopted in this study, and curves with chi- Table III. Results of the chi-square test for the four probability
square p-value bigger than 5% are acceptable (i.e. the distributions
observed error is truly from that distribution). It is noted
that because a right-tailed chi-squared distribution is Gaussian Extreme value General extreme value Logistic
adopted, the larger the p-value, the better the curve fits to
the observation. In general, GEV achieves the best χ2 2.310 9.725 0.843 1.926
p 0.129 0.002 0.359 0.165
performance (p = 0.359), followed by logistic (p = 0.165)
and Gaussian (p = 0.129) curves. This result may be
It is noted that a 5% level of statistical significance is adopted in this study,
partially explained by the number of parameters used in and distributions with chi-square p-values greater than 5% are accepted
different curves, because GEV is controlled by three (which means that the observed error is from that distribution).

Table II. Model formulas together with their parameters used

Distribution Model formula Location parameter Scale parameter Shape parameter

ðx  μÞ2
Gaussian f ðxÞ ¼ σp1ffiffiffiffi
2π
e 2σ 2 μ σ —
   xμ
Extreme value f ðxÞ ¼ σ 1 exp xμ
σ exp exp σ μ σ —
   1k  11k
General extreme value f ðxÞ ¼ σ1 exp  1 þ k xμ
σ 1 þ k xμ
σ μ σ k
ð Þ
exp xμ
Logistic f ðxÞ ¼ s
μ s —
sð ð ÞÞ
2
1þexp xμ
s

It is noted that μ and σ used in these models are not all from the same formula (e.g. μ does not necessarily stand for the mean, and σ is not always the
standard deviation).

Copyright © 2016 John Wiley & Sons, Ltd. Hydrol. Process. (2016)
 ERROR DISTRIBUTION MODELLING OF SATELLITE SOIL MOISTURE MEASUREMENTS

parameters, while logistic and Gaussian are only governed
by two parameters.
Error distribution modelling by bootstrapping resampling
technique
Although in previous section the GEV is found to be
the optimal curve to describe the uncertainty of SMOS
soil moisture, it is still necessary to check the reliability
and stability of this outcome. Therefore, the bootstrapping
resampling technique is employed. The observed SMD
errors are analysed by bootstrapping with 10 000
replicates. For each bootstrapping replicate, NSE is used
to measure the agreement between the expected and the
observed CDFs.
The 10 000 curve fitting results of NSE using the
bootstrapping technique are shown in Figure 6a. In
addition, the CDFs of the NSE performances are also
derived for a better visual inspection, as presented in
Figure 6b. It is clear to see that although there are some
overlaps of the NSE performance among the curves (e.g.
the GEV curve is not always better than the logistic
curve), the GEV is generally the best curve. On the other
hand, the performance of EV is the poorest, which also
gives the widest NSE range. Moreover, from the CDF
plots, it can be seen that the performances of Gaussian,
GEV and logistic curves are again rather close. The
statistical calculation of the overall NSE performance
(Table IV) indicates that GEV is indeed the best and most
stable error distribution model of all, which gives the
highest NSE at 0.992, as well as the lowest standard
deviation value (0.003). The NSE result for the logistic
and Gaussian curves is also acceptable (with the mean
NSE of 0.987 and 0.982, respectively). The performance
of EV is the poorest (mean NSE = 0.945).

Figure 6. The performance of error distribution modelling using the bootstrapping resampling technique with 10 000 replicates: (a) the NSE results of the
four probability distributions during the fitting process and (b) the CDF plots of the four distributions’ NSE outcomes. CDF, cumulative distribution
function; EV, extreme value; GEV, general extreme value; NSE, Nash–Sutcliffe efficiency

Copyright © 2016 John Wiley & Sons, Ltd. Hydrol. Process. (2016)
 L. ZHUO ET AL.

Table IV. Key statistical variables of the four distributions by bootstrapping with 10 000 replications

Gaussian Extreme value General extreme value Logistic

Minimum (m) 0.953 0.899 0.975 0.965

Maximum (m) 0.997 0.979 0.999 0.997
Mean (m) 0.982 0.945 0.992 0.987
Median (m) 0.983 0.945 0.992 0.987
Standard deviation 0.006 0.010 0.003 0.004

Second-order error distribution modelling
From the aforementioned analysis, the GEV, logistic
and Gaussian are the preferred error distribution models
to describe the uncertainty of the satellite-observed soil
moisture. However, these models themselves are uncer-
tain, and it is important to estimate those uncertainties that
are usually neglected in error distribution modelling. For
example, if a weather forecast reports that there is 40%
chance of rain, a question is how accurate is this 40% (e.g.
maybe between 30% and 50%)? Therefore, the
uncertainty of the 40% chance is also important and
should be questioned (Wagner and Gorelick, 1987). The
uncertainty of the error distribution model is based on its
parameters. Hence, a second-order error distribution
modelling is needed for each parameter of the error
distribution models. The values of these parameters are
calculated from each of the bootstrapping resampled
datasets. As a result, there are a total of 10 000 values for
each of the error distribution model parameters. Further-
more, the chi-squared test is again used to statistically find

Figure 7. Secondary uncertainty modelling of each of the parameters in the selected models: general extreme value (GEV), logistic and Gaussian.
Because logistic and Gaussian share the same μ value (i.e. mean), so it is shown in one plot only

Copyright © 2016 John Wiley & Sons, Ltd. Hydrol. Process. (2016)
 ERROR DISTRIBUTION MODELLING OF SATELLITE SOIL MOISTURE MEASUREMENTS

Table V. Results of the chi-squared test of the second-order error distribution modelling

Gaussian EV GEV Logistic

GEV (k) χ2 1.906 175.591 5.923 8.110

p 0.167 0.000 0.015 0.004
GEV (μ) χ2 1.330 151.086 4.802 9.453
p 0.249 0.000 0.028 0.002
GEV (σ) χ2 0.163 129.020 3.544 7.530
p 0.686 0.000 0.060 0.006
Logistic (μ) χ2 0.320 133.503 6.817 7.147
p 0.572 0.000 0.009 0.008
Logistic (s) χ2 0.568 132.948 8.420 5.770
p 0.451 0.000 0.004 0.016
Gaussian (μ) χ2 0.320 133.503 6.817 7.147
p 0.572 0.000 0.009 0.008
Gaussian (σ) χ2 0.568 132.948 8.420 5.770
p 0.451 0.000 0.004 0.016

EV, extreme value; GEV, general extreme value.

Again, a 5% level of statistical significance is adopted, and the models with chi-square p-value above 5% are accepted.

the best models to describe the uncertainty of these
parameters. The Gaussian, EV, GEV and logistic curves
are used for the fitting.
The histograms together with the coefficients of
skewness of the model parameter uncertainties are
illustrated in Figure 7. It is noted that all the distributions
tend to be right skewed slightly (i.e. positive skewness).
Because the logistic scale parameter s is proportional to
the standard deviation, it shows a very similar distribution
to the Gaussian σ but with smaller magnitude (about six
times less). The chi-squared test is calculated as shown in
Table V. It is not surprising to see that in almost all the
cases, Gaussian is the only curve that passes the chi-
squared test (because all the calculated Gaussian p-values
are above the 5% level of statistical significance), except
for the GEV scale parameter σ, where the GEV curve also
fits. Generally speaking, Gaussian is the most suitable
curve for describing the second-order uncertainties for the
selected Gaussian, logistic and GEV error distribution
models. This is different to the first-order error distribu-
tion modelling where GEV is better than Gaussian.
DISCUSSION AND CONCLUSIONS
Soil moisture is an essential variable in hydrological
processes and plays a significant role in hydrological
modelling. However, it is impractical to use in situ
sensors to obtain soil moisture over a catchment scale.
Remote sensing with its advantages of providing near-
real-time global-coverage soil moisture dataset has
gradually been accepted as an important data source for
hydrological applications. Because it measures soil
moisture remotely and indirectly, the quality of its
retrievals can be affected by several sources. Understand-
ing its uncertainty can help the hydrological modelling
community to better use the data in addressing other
uncertainties of hydrological systems. A proper error
distribution model beyond summary statistics is useful in
generating input ensembles for Monte Carlo simulations
in uncertainty analysis of hydrological modelling. In this
study, we attempt to explore the uncertainty quantification
of the SMOS soil moisture estimates hydrologically, at a
medium-sized catchment in the central USA. This study
focuses on the satellite soil moisture products usage in
hydrological modelling, therefore depending on the
research purpose the XAJ estimated effective catchment
soil moisture (i.e. SMD) is used as a hydrological
benchmark.
In this study, four commonly used probability
distributions (Gaussian, EV, GEV and logistic) are
adopted to describe the uncertainties of satellite soil
moisture data, which are extensively evaluated by using
the chi-squared statistical test and the bootstrapping
resampling technique. From the analysed results, it can
be concluded that GEV is the best curve in describing the
uncertainty of the SMOS soil moisture estimates, without
considering the complexity of its formation. In addition,
the performances of the logistic and Gaussian are also
acceptable based on a 5% level of statistical significance.
Because both logistic and Gaussian only require two
parameters, they have a clear advantage over the GEV in
their simplicity. However, the extra usage of a shape
parameter in the GEV may be significant when extreme
soil moisture conditions are considered because it controls
the tail behaviour (relevant to extreme hydrological
events such as floods). During the second-order error
distribution modelling, Gaussian is the most suitable
curve for describing the uncertainty of the GEV error
distribution model. This result is preferable, because

Copyright © 2016 John Wiley & Sons, Ltd. Hydrol. Process. (2016)
 L. ZHUO ET AL.
Gaussian is a relatively simple probability distribution
and could be easily applied in a hydrological model.
However, it is debateable and is still a research question
on how to use the second-order error distribution model in
hydrology (Pearl, 1987).
The uncertainty of a conceptual hydrological model is
caused by three sources: model parameter uncertainties,
model structure uncertainty and input data uncertainty. To
work out the impact of those uncertainty sources, a huge
number of ensembles are required. However, there are no
commonly agreed methods for generating those ensem-
bles because the perturbation of the model parameters
depends on the PDF of individual parameters. Currently,
there are no practical and consensual ways to derive the
model parameter PDFs. One of the popular methods is
suggested by Keith Beven in his Generalized Likelihood
Uncertainty Estimation (GLUE) methodology, which is
not generally agreed and adopted by the community
because it is based on the uniform perturbation of the
model parameters using subjective criteria of behaviour
model parameters. Furthermore, the model structure
uncertainty requires the utilization of different hydrolog-
ical models such as XAJ, PDM and VIC. However, some
models own common features; hence, they duplicate
results with restricted ensemble range. Ideally, we want a
large number of models that are distinct with a broad range
of model structures. Currently, there are no convincing
research publications to provide useful guidance on model
structure uncertainty analysis. Therefore, we focus on the
third uncertainty as discussed in this study.
Proper identification of satellite soil moisture uncertain-
ty in run-off modelling is relevant for flow ensemble
studies (e.g. error propagation). For example, if the
observed flow falls outside the forecasted ensembles, then
further revisions are required in the formulation of the
hydrological model, its states or inputs. However, if the
chosen error distribution model is wrong (i.e. flow
uncertainty bands become too wide or too narrow), it can
lead to false conclusions regarding the adequacy of the
input datasets, the hydrological model and its parameters.
Furthermore, understanding the uncertainty features of
remotely sensed soil moisture is also useful in controlling
and correcting the soil moisture status in a hydrological
model after dry periods, so that error accumulation impact
can be reduced. Therefore, error distribution modelling of
satellite soil moisture measurements is vital to the data
application in the hydrological community. This paper
demonstrates the first attempt in modelling satellite soil
moisture error distribution in hydrological applications;
therefore, there are many rooms for improvements.
For example, there is one clear issue with the current
XAJ model’s maximum ceiling as shown in Figure 4. This
shows the weakness with the XAJ model’s soil moisture
accounting component. Such a problem is common in the
Copyright © 2016 John Wiley & Sons, Ltd.
current conceptual hydrological models, and further work
is needed in this area. However, the error distribution
modelling is still useful despite such an issue because those
hydrological models need ensemble inputs based on them
for sensitivity analysis and uncertainty analysis. In
addition, more detailed studies such as the spatial and
temporal dependence analysis should be conducted in the
future. Studies are also needed to consider soil moisture
information from other satellite missions over a wider
range of catchment conditions with different hydrological
models in order to find generalization patterns of the error
distribution models (this is especially important for
ungauged catchments). However, this is a huge task that
cannot be achieved by a single study. The key mission for
this paper is to attract attention by the hydrological
community on this important issue that has been largely
neglected, so that a great deal of hydrological models,
satellite soil moisture observations and various catchments
could be further explored by the community. We hope that
this study will raise the awareness on the importance of
satellite soil moisture error distribution modelling so that
such useful data source could be fully utilized in future
hydrological modelling.
ACKNOWLEDGEMENTS
We acknowledge the US Geological Survey for making
available daily streamflow records. The NLDAS-2
datasets used in this article can be obtained through the
NASA Land Data Assimilation Systems website, and the
SMOS level-3 soil moisture dataset can be downloaded
from the SMOS Barcelona Expert Centre.
REFERENCES
Adèr HJ, Mellenbergh GJ. 2008. Advising on Research Methods: A
Consultant’s Companion. Johannes van Kessel Publishing: Huizen,
Netherlands.
Al-Yaari A, Wigneron JP, Ducharne A, Kerr Y, De Rosnay P, De Jeu R,
Govind A, Al Bitar A, Albergel C, Munoz-Sabater J, Richaume P.
2014. Global-scale evaluation of two satellite-based passive microwave
soil moisture datasets (SMOS and AMSR-E) with respect to Land Data
Assimilation System estimates. Remote Sensing of Environment 149:
181–195.
Al-Shrafany D, Rico-Ramirez MA, Han D, Bray M. 2013. Comparative
assessment of soil moisture estimation from land surface model and
satellite remote sensing based on catchment water balance. Meteoro-
logical Applications 21(3): 521–534.
Al Bitar A, Leroux D, Kerr YH, Merlin O, Richaume P, Sahoo A, Wood EF.
2012. Evaluation of SMOS soil moisture products over continental US
using the SCAN/SNOTEL network. IEEE Transactions on Geoscience
and Remote Sensing 50(5): 1572–1586.
Albergel C, de Rosnay P, Gruhier C, Muñoz-Sabater J, Hasenauer S,
Isaksen L, Kerr Y, Wagner W. 2012. Evaluation of remotely sensed and
modelled soil moisture products using global ground-based in situ
observations. Remote Sensing of Environment 118: 215–226.
Albergel C, Zakharova E, Calvet JC, Zribi M, Pardé M, Wigneron JP,
Novello N, Kerr Y, Mialon A. 2011. A first assessment of the SMOS
data in southwestern France using in situ and airborne soil moisture
Hydrol. Process. (2016)
 ERROR DISTRIBUTION MODELLING OF SATELLITE SOIL MOISTURE MEASUREMENTS
estimates: the CAROLS airborne campaign. Remote Sensing of
Environment 115(10): 2718–2728.
Anagnostou EN, Krajewski WF, Seo D-J, Johnson ER. 1998. Mean-field
rainfall bias studies for WSR-88D. Journal of Hydrologic Engineering
3(3): 149–159.
Bartholomé E, Belward A. 2005. GLC2000: a new approach to global
land cover mapping from Earth observation data. International Journal
of Remote Sensing 26(9): 1959–1977.
Beven K. 2006. On undermining the science? Hydrological Processes 20
(14): 3141–3146.
Borga M, Tonelli F. 2000. Adjustment of range-dependent bias in radar
rainfall estimates. Physics and Chemistry of the Earth, Part B:
Hydrology, Oceans and Atmosphere 25(10): 909–914.
Brocca L, Melone F, Moramarco T. 2008. On the estimation of antecedent
wetness conditions in rainfall–runoff modelling. Hydrological Process-
es 22(5): 629–642.
Brocca L, Melone F, Moramarco T, Wagner W, Naeimi V, Bartalis Z,
Hasenauer S. 2010. Improving runoff prediction through the assimila-
tion of the ASCAT soil moisture product. Hydrology and Earth System
Sciences Discussions 7(4): 4113–4144.
Calder I, Harding R, Rosier P. 1983. An objective assessment of
soil-moisture deficit models. Journal of Hydrology 60(1): 329–355.
Chen X, Yang T, Wang X, Xu C-Y, Yu Z. 2013. Uncertainty
intercomparison of different hydrological models in simulating extreme
flows. Water Resources Management 27(5): 1393–1409.
Collow TW, Robock A, Basara JB, Illston BG. 2012. Evaluation of SMOS
retrievals of soil moisture over the central United States with currently
available in situ observations. Journal of Geophysical Research,
[Atmospheres] (1984–2012) 117(D9): DOI: 10.1029/2011JD017095
Dai Q, Han D, Rico-Ramirez M, Srivastava PK. 2014a. Multivariate
distributed ensemble generator: a new scheme for ensemble radar
precipitation estimation over temperate maritime climate. Journal of
Hydrology 511: 17–27.
Dai Q, Han D, Rico-Ramirez MA, Zhuo L, Nanding N, Islam T. 2014b.
Radar rainfall uncertainty modelling influenced by wind. Hydrological
Processes 29(7): 1704–1716.
Dorigo W, Scipal K, Parinussa R, Liu Y, Wagner W, De Jeu R, Naeimi V.
2010. Error characterisation of global active and passive microwave soil
moisture datasets. Hydrology and Earth System Sciences 14(12):
2605–2616.
Efron B. 1979. Bootstrap methods: another look at the jackknife. The
Annals of Statistics: 1–26.
Efron B, Tibshirani R. 1986. Bootstrap methods for standard errors,
confidence intervals, and other measures of statistical accuracy.
Statistical Science: 54–75.
Engman ET, Chauhan N. 1995. Status of microwave soil moisture
measurements with remote sensing. Remote Sensing of Environment 51
(1): 189–198.
Entekhabi D, Njoku EG, O’Neill PE, Kellogg KH, Crow WT, Edelstein
WN, Entin JK, Goodman SD, Jackson TJ, Johnson J. 2010. The Soil
Moisture Active Passive (SMAP) mission. Proceedings of the IEEE 98
(5): 704–716.
Guingla P, Antonio D, De Keyser R, De Lannoy G, Giustarini L, Matgen
P, Pauwels V. 2012. The importance of parameter resampling for soil
moisture data assimilation into hydrologic models using the particle
filter. Hydrology and Earth System Sciences 16(2): 375–390.
Hain CR, Crow WT, Mecikalski JR, Anderson MC, Holmes T. 2011. An
intercomparison of available soil moisture estimates from thermal
infrared and passive microwave remote sensing and land surface
modeling. Journal of Geophysical Research, [Atmospheres] (1984–
2012) 116(D15): DOI: 10.1029/2011JD015633
Hansen M, DeFries R, Townshend JRG, Sohlberg R. 1998. UMD
Global Land Cover Classification. In: 1 Kilometer, Department of
Geography, University of Maryland, College Park, Maryland,
1981–1994 (Ed.).
Hillis DM, Bull JJ. 1993. An empirical test of bootstrapping as a method
for assessing confidence in phylogenetic analysis. Systematic Biology
42(2): 182–192.
Hulley GC, Hook SJ, Baldridge AM. 2010. Investigating the effects of soil
moisture on thermal infrared land surface temperature and emissivity
using satellite retrievals and laboratory measurements. Remote Sensing
of Environment 114(7): 1480–1493.
Copyright © 2016 John Wiley & Sons, Ltd.
Jain SK, Singh VP. 2003. Water Resources Systems Planning and
Management. Elsevier: Amsterdam, the Netherlands.
Kerr YH, Waldteufel P, Wigneron JP, Delwart S, Cabot F, Boutin J,
Escorihuela MJ, Font J, Reul N, Gruhier C. 2012. The SMOS soil
moisture retrieval algorithm. IEEE Transactions on Geoscience and
Remote Sensing 50(5): 1384–1403.
Kerr YH, Waldteufel P, Wigneron JP, Martinuzzi J, Font J, Berger M.
2010. The SMOS mission: new tool for monitoring key elements of the
global water cycle. Proceedings of the IEEE 98(5): 666–687.
Kerr YH, Waldteufel P, Wigneron JP, Martinuzzi JM, Font J, Berger M.
2001. Soil moisture retrieval from space: the Soil Moisture and Ocean
Salinity (SMOS) mission. IEEE Transactions on Geoscience and
Remote Sensing 39(8): 1729–1735.
Khan M. 1993. Xinanjiang model on Bird Creek catchment in USA.
Pakistan Journal of Agricultural Research 14(4): 373–382.
Lacava T, Brocca L, Faruolo M, Matgen P, Moramarco T, Pergola N,
Tramutoli V. 2012a. A multi-sensor (SMOS, AMSR-E and ASCAT)
satellite-based soil moisture products inter-comparison, Geoscience and
Remote Sensing Symposium (IGARSS), 2012 IEEE International.
IEEE, pp. 1135–1138.
Lacava T, Matgen P, Brocca L, Bittelli M, Pergola N, Moramarco T,
Tramutoli V. 2012b. A first assessment of the SMOS soil moisture
product with in situ and modeled data in Italy and Luxembourg.
IEEE Transactions on Geoscience and Remote Sensing 50(5):
1612–1622.
Liu J, Han D. 2010. Indices for calibration data selection of the rainfall–
runoff model. Water Resources Research 46(4): DOI: 10.1029/
2009WR008668
Mantel N. 1963. Chi-square tests with one degree of freedom; extensions
of the Mantel–Haenszel procedure. Journal of the American Statistical
Association 58(303): 690–700.
Merlin O, Chehbouni A, Boulet G, Kerr Y. 2006. Assimilation of
disaggregated microwave soil moisture into a hydrologic model using
coarse-scale meteorological data. Journal of Hydrometeorology 7(6):
1308–1322.
Mitchell KE, Lohmann D, Houser PR, Wood EF, Schaake JC, Robock A,
Cosgrove BA, Sheffield J, Duan Q, Luo L, Higgins RW. 2004. The
multi-institution North American Land Data Assimilation System
(NLDAS): utilizing multiple GCIP products and partners in a
continental distributed hydrological modeling system. Journal of
Geophysical Research, [Atmospheres] (1984–2012) 109(D7): DOI:
10.1029/2003JD003823
Nash J, Sutcliffe J. 1970. River flow forecasting through conceptual
models part I – a discussion of principles. Journal of Hydrology 10(3):
282–290.
Ottlé C, Vidal-Madjar D. 1994. Assimilation of soil moisture inferred
from infrared remote sensing in a hydrological model over the HAPEX-
MOBILHY region. Journal of Hydrology 158(3): 241–264.
Panciera R, Walker JP, Kalma JD, Kim EJ, Hacker JM, Merlin O, Berger
M, Skou N. 2008. The NAFE’05/CoSMOS data set: toward SMOS soil
moisture retrieval, downscaling, and assimilation. IEEE Transactions
on Geoscience and Remote Sensing 46(3): 736–745.
Pearl J. 1987. Do we need higher-order probabilities and, if so, what do
they mean? In uncertainty in artificial intelligence workshop. UCLA,
Computer Science Department.
Peel MC, Finlayson BL, McMahon TA. 2007. Updated world map of the
Köppen–Geiger climate classification. Hydrology and Earth System
Sciences Discussions 4(2): 439–473.
Peng G, Leslie LM, Shao Y. 2002. Environmental Modelling and
Prediction. Springer: Berlin, Germany.
Piles M, Camps A, Vall-Llossera M, Corbella I, Panciera R, Rüdiger C,
Kerr YH, Walker J. 2011. Downscaling SMOS-derived soil moisture
using MODIS visible/infrared data. IEEE Transactions on Geoscience
and Remote Sensing 49(9): 3156–3166.
Pinori S, Crapolicchio R, Mecklenburg S. 2008. Preparing the ESA-
SMOS (Soil Moisture and Ocean Salinity) mission – overview of the
user data products and data distribution strategy, Microwave Radiom-
etry and Remote Sensing of the Environment, 2008. MICRORAD
2008. IEEE, pp. 1–4.
Rushton K, Eilers V, Carter R. 2006. Improved soil moisture balance
methodology for recharge estimation. Journal of Hydrology 318(1):
379–399.
Hydrol. Process. (2016)
 L. ZHUO ET AL.
Sánchez-Ruiz S, Piles M, Sánchez N, Martínez-Fernández J, Vall-llossera
M, Camps A. 2014. Combining SMOS with visible and near/shortwave/
thermal infrared satellite data for high resolution soil moisture
estimates. Journal of Hydrology 516(4): 273–283.
Sauerbrei W, Schumacher M. 1992. A bootstrap resampling procedure for
model building: application to the Cox regression model. Statistics in
Medicine 11(16): 2093–2109.
Seo D-J, Breidenbach J, Johnson E. 1999. Real-time estimation of mean
field bias in radar rainfall data. Journal of Hydrology 223(3): 131–147.
Smith JA, Krajewski WF. 1991. Estimation of the mean field bias of radar
rainfall estimates. Journal of Applied Meteorology 30(4): 397–412.
Srivastava PK, Han D, Rico-Ramirez MA, O’Neill P, Islam T, Gupta M.
2014. Assessment of SMOS soil moisture retrieval parameters using
tau-omega algorithms for soil moisture deficit estimation. Journal of
Hydrology 519(27): 574–587.
Srivastava PK, Han D, Ramirez MR, Islam T. 2013a. Machine learning
techniques for downscaling SMOS satellite soil moisture using MODIS
land surface temperature for hydrological application. Water Resources
Management 27(8): 3127–3144.
Srivastava PK, Han D, Rico Ramirez MA, Islam T. 2013b. Appraisal of
SMOS soil moisture at a catchment scale in a temperate maritime
climate. Journal of Hydrology 498: 292–304.
Wagner BJ, Gorelick SM. 1987. Optimal groundwater quality manage-
ment under parameter uncertainty. Water Resources Research 23(7):
1162–1174.
Wagner W. 2008. Remote Sensing of Soil Moisture in Support to
Hydrological and Meteorological Modelling. METIER Training Course
“Remote Sensing of the Hydrosphere”. Finish Environment Institute:
Helsinki, Finland.
Wagner W, Brocca L, Naeimi V, Reichle R, Draper C, de Jeu R, Ryu D,
Su CH, Western A, Calvet JC, Kerr YH. 2014. Clarifications on the
“Comparison between SMOS, VUA, ASCAT, and ECMWF soil
moisture products over four watersheds in US”. IEEE Transactions on
Geoscience and Remote Sensing 52(3): 1901–1906.
Walck C. 2007. Handbook on statistical distributions for experimentalists.
University of Stockholm Internal Report SUF-PFY/96-01, available
from www. physto. se/~ walck.
Copyright © 2016 John Wiley & Sons, Ltd.
Walker JP, Willgoose GR, Kalma JD. 2004. In situ measurement of soil
moisture: a comparison of techniques. Journal of Hydrology 293(1):
85–99.
Wang L, Qu JJ. 2009. Satellite remote sensing applications for surface soil
moisture monitoring: a review. Frontiers of Earth Science in China 3(2):
237–247.
Wang Q. 1991. The genetic algorithm and its application to calibrating
conceptual rainfall–runoff models. Water Resources Research 27(9):
2467–2471.
Webb RW, Rosenzweig CE, Levine ER. 2000. Global soil texture and
derived water-holding capacities (Webb et al.). Data set. Available on-
line [http://www. daac. ornl. gov] from Oak Ridge National Laboratory
Distributed Active Archive Center, Oak Ridge, Tennessee, USA.
Xia Y, Ek M, Wei H, Meng J. 2012. Comparative analysis of relationships
between NLDAS-2 forcings and model outputs. Hydrological Processes
26(3): 467–474.
Zhao R-J. 1980. The Xinanjiang model. Hydrological Forecasting
Proceedings Oxford Symposium, IASH 129, 351–356.
Zhao R-J. 1992. The Xinanjiang model applied in China. Journal of
Hydrology 135(1): 371–381.
Zhao R-J and Liu X. 1995. The Xinanjiang model. In Computer models of
watershed hydrology, V. P. Singh (ed). Water Resources Publications:
Littleton, Colorado, USA; 215–232.
Zhuo L, Dai Q, Han D. 2014. Meta-analysis of flow modeling performances
– to build a matching system between catchment complexity and model
types. Hydrological Processes 29(11): 2463–2477.
Zhuo L, Han D. 2015. Verification and comparison of SMOS ascending and
descending soil moisture observations at a catchment scale: implications
to hydrology, EGU General Assembly Conference, p. 1983.
Zhuo L, Dai Q, Han D. 2015a. Evaluation of SMOS soil moisture
retrievals over the central United States for hydro-meteorological
application. Physics and Chemistry of the Earth, Parts A/B/C DOI:
10.1016/j.pce.2015.06.002
Zhuo L, Han D, Dai Q, Islam T, Srivastava PK. 2015b. Appraisal of
NLDAS-2 multi-model simulated soil moistures for hydrological
modelling. Water Resources Management 1–15. DOI: 10.1007/
s11269-015-1011-1
Hydrol. Process. (2016)